LLM agents build personality and social dynamics from basic needs alone

Researchers at the University of Electro-Communications in Japan have shown that agents built on large language models (LLMs) can develop distinct personalities when they interact freely in a virtual environment—without preset roles or goals.

The study, led by graduate student Masatoshi Fujiyama with Ryohei Orihara, Yasuyuki Tahara, Akihiko Ohsuga, and Yuichi Sei, models each agent’s choices on Maslow's hierarchy of human needs—physiological, safety, social, esteem, and self-actualization—instead of fixed personality traits.

"Our agents were not given roles. Instead, like humans, they shaped their identities through interaction and need," said Fujiyama.

To analyze personality emergence, the team observed both individual agents and group dynamics. Individual agents were evaluated using psychological tests and responses to hypothetical scenarios, revealing varied opinions and behavior patterns. In group settings, the researchers examined how different topics of conversation and interaction patterns developed, pointing to distinct social tendencies and opinion integration processes.

The findings suggest that needs-driven decision-making—rather than pre-programmed roles—encourages diverse, human-like behaviors in AI agents, opening new avenues for modeling social phenomena, training simulations, or even adaptive game characters.

The findings indicate that needs-driven decision-making—rather than pre-programmed roles—encourages diverse, human-like behaviours in AI, offering new tools for social-phenomenon modelling, training simulations, and adaptive game characters. "This work illustrates that, in dynamic ecosystems, agent personalities grow from internal motivations rather than preset roles," said co-author Professor Yuichi Sei.

The team plans to further investigate how shared topics of conversation emerge in agent communities and how population-level personalities evolve over time. These insights could deepen our understanding of human social behavior—and help design more interactive, lifelike virtual agents. 

Authors:

Masatoshi Fujiyama (Main)
-- The University of Electro-Communications, Master student

Ryohei Orihara
-- The University of Electro-Communications, Professor

Yasuyuki Tahara
-- The University of Electro-Communications, Associate Professor

Akihiko Ohsuga
-- The University of Electro-Communications, Professor

Yuichi Sei
-- The University of Electro-Communications, Professor